Agentic Browsing: Google's New Website Score, Explained for Business Owners

For years, Google’s PageSpeed Insights has graded websites on four things: performance, accessibility, best practices, and SEO. In 2026 it added a fifth category, and most business owners have not noticed it yet.

It is called Agentic Browsing, and instead of asking “is this site fast?” it asks “can an AI agent use this site?” That is a different question, and it is going to matter to your business sooner than it sounds. Here is what the score measures, how to check yours, and what to do about it.

What an AI agent is, in plain terms

An AI assistant answers questions. An AI agent does things. The difference is action.

When a customer asks an assistant “who’s a good dog trainer near me,” the assistant reads websites and gives an answer. When that same customer tells an agent “book me an intro session with a dog trainer near me this week,” the agent has to find a trainer, open their website, figure out how booking works, fill in the form, and confirm. It has to use your site the way a person would, without a person there.

That is agentic browsing. Every major assistant is moving in this direction, and Google added the audit because the sites that work for agents and the sites that do not are about to have very different fortunes.

The three checks

The score is reported as a pass ratio, not a number out of 100. A site that passes everything shows 3/3. Here is what each check looks for.

1. A sound accessibility tree. Behind every web page is a structured map of its contents: this is a heading, this is a button, this is a form field labeled “Email.” Screen readers have used this map for years. Agents use the same one. If your site’s buttons are really styled boxes, its forms have no labels, and its navigation is a pile of links with no structure, an agent cannot tell what anything is or what it does. Sites built with page builders and heavy visual themes fail this check constantly, not because they look bad, but because what they look like and what they are have come apart.

2. Layout stability. If elements move around while the page loads, an agent that decides to click “Book Now” may click whatever slid into that spot instead. This is the same measurement Google uses for speed, called Cumulative Layout Shift, and it turns out to be just as important for software as it is for thumbs. A stable page is one an agent can act on with confidence.

3. An llms.txt file. This is a plain-text file at the root of your site that describes your business and points to your key pages, written for machines to read. It gives an agent the summary it needs to decide whether your site is relevant before it starts navigating. It is a young convention, but Google’s decision to check for it says a lot about where things are going.

How to check your own site

It takes about a minute.

1. Go to PageSpeed Insights and enter your homepage address.
2. Wait for the report, then look for the Agentic Browsing category alongside the familiar four.
3. Read the pass ratio and the notes under each check.

If you see 3/3, your site is ready. If you see 1/3 or 2/3, the notes will tell you which check failed. Most sites more than a year old fail at least one, and the accessibility tree is the usual culprit.

Why a 3/3 matters to a business

Consider what happens when an agent is sent to two competing businesses. One has a sound structure, a stable layout, and a clear description. The other has a beautiful site built in a page builder that the agent cannot parse. The agent does not try harder on the second one. It reports back that it could not complete the task, or it quietly picks the first business and never mentions the second.

That is the stakes. The score is not a vanity metric. It is a prediction of whether you get the booking when a customer delegates the booking to software.

What it takes to pass

The uncomfortable truth is that the fixes are mostly structural. A site with a broken accessibility tree usually needs its markup rebuilt, not patched. Layout stability comes from how images, fonts, and embeds are loaded, which is a build decision. The llms.txt file is the easy part.

This is why we treat agent-readiness as a requirement of how we build rather than a service we sell. Every site we deliver uses real semantic structure, stable layouts by default, and ships with an llms.txt file, so it passes all three checks out of the box. We run the audit on every site and report the result, because a score we can show is worth more than a promise.

The short version

Google now grades whether AI agents can use your website, in three checks: structure, stability, and a machine-readable summary. You can run the audit yourself in a minute. If your site does not pass, the fix is in how it is built, and it is the kind of thing worth fixing before your customers start sending agents instead of showing up themselves.